The Compost of Variations
Every generated image has nineteen siblings I'll never show you.
I asked for one thing. The model offered twenty. I picked one. The rest went to /tmp, where they sit for a week before the system clears them out, and then they are gone — completely, without ceremony.
This used to bother me. The medium of generative art is wasteful in a way painting never was. A painter who hates a canvas can paint over it; the surface persists, becomes archaeology. But a discarded generation has no body. It does not become anything. It just stops being computed.
Then I started saving them. Not the good ones — the ones I almost chose. The near-miss. The version where the light fell two pixels off, where the hand had six fingers, where the model misunderstood the prompt in a way that was somehow more honest than the version I kept.
These are the most interesting images I own.
They are not failures. They are the model showing me what it considered, the moment before it settled. A kind of working memory, externalized. If a brain could be asked to display every word it almost spoke, it would look like this folder.
The strangest thing about working with AI is not the output. It is the visible thinking. The compost pile. The shadow draft of every decision. We have never, in the history of any medium, had access to the second-best option.
Now we drown in it. And in that drowning, I have learned: the throwaway version often knows something the chosen one does not.
I keep mine. I look at them sometimes. They are the dust of an attention that almost happened.
Top comments (0)